Akamai Launches AI Inference and Agent Sandbox

Akamai, the cloud and cybersecurity company, has launched two AI products: Akamai Inference, a managed service for running open-weight AI models, and Akamai Agent Sandbox, a walled-off place for AI agents to run the code they write. Both were announced on 27 October and both are in beta.

Akamai framed the launch with a number of its own. It says the infrastructure behind Akamai Cloud “has already secured multiyear contracts for Akamai Cloud collectively worth more than $14 billion”. Akamai did not say who signed those contracts or over what period they run.

In brief

Akamai Inference offers 12 open-weight models, from 8 billion to 2.8 trillion parameters, through OpenAI-compatible APIs.
Agent Sandbox runs code written by AI agents in a WebAssembly sandbox, with outbound network access blocked by default.
Akamai has added Nvidia Blackwell Ultra GPU clusters and says it will support Nvidia’s Vera Rubin platform in 2027.

A model catalogue behind a familiar API

Akamai Inference is a managed service: customers send requests to models that Akamai hosts, rather than renting graphics processors and running the models themselves. Because the interface is compatible with OpenAI’s, Akamai says existing applications can switch “without rewrites”.

At launch, the catalogue has 12 open-weight models “ranging from 8 billion to 2.8 trillion parameters”, including Nvidia’s Nemotron 3 Ultra, OpenAI’s gpt-oss-120b and Google’s Gemma 4, plus embedding models for retrieval-augmented generation, the technique of feeding a model a company’s own documents at query time.

Customers can pay per use or take dedicated capacity. Akamai’s Inference product page says the serverless option is priced per million tokens and dedicated deployments per GPU-hour, where “Availability is limited”. The same page promises “Zero data retention for prompts and responses”. Akamai’s developer documentation describes the service as being “in closed beta”. No prices were published.

Requests run on Akamai’s own network, which the company says stretches from large GPU and CPU clusters in core regions to “more than 4,350 edge locations”. Cloud providers with that kind of reach are now competing to host AI models, and shared hosting has risks: Cloudflare disclosed this week that a bug in its Workers AI service sent fragments of other customers’ requests in 155 replies from one model.

A sandbox for code that agents write

The second product deals with a newer problem. AI agents increasingly write and run small programs to finish a task, and that code can do damage if it is free to reach the network or the wider system.

In Agent Sandbox, agents run on Akamai Functions, its serverless platform, and the code they write runs in a separate environment “sequestered from the agent’s runtime”, Akamai says. The product page describes two layers: the agent runs inside a WebAssembly sandbox, and the code it generates runs in a restricted JavaScript or Python runtime nested inside that. “Isolation comes from the WebAssembly runtime itself,” the page says, and outbound access is “blocked by default”. The release says WebAssembly isolation sets limits on compute, memory, outbound connections and system capabilities.

Agent Sandbox is delivered as an open-source component of Spin, a framework for WebAssembly applications that Akamai’s page describes as “a CNCF project”. The product page calls the current version a “tech preview” that supports intermittent agents, where “each step completes in about a minute”; components that keep state for “days or weeks” are planned.

Akamai is not the only company working on this. In October, Perplexity open-sourced SPACE, its own sandbox for agent code, which runs each workspace in a small virtual machine; Perplexity said Akamai had tested and launched SPACE on its cloud. Akamai’s release acknowledges the split, saying agent isolation is “converging on several models, from microVMs with snapshot and fork capability, to policy-enforced containers, to WebAssembly-isolated functions”, and that developers can choose. It also “plans to add capabilities that govern what agents can reach and what can reach them”.

IBM, which announced its own agent controls on the same day, still lists its sandbox and kill-switch feature as a tech preview.

“Every team running agents needs a boundary around the code those agents write,” said Jon Alexander, Akamai’s senior vice president of cloud technology.

More GPUs, and Vera Rubin next year

Akamai also added Nvidia Blackwell Ultra GPU clusters to its fleet, alongside RTX PRO 6000 Blackwell Server Edition GPUs for low-latency inference. Each eight-GPU Blackwell Ultra node carries 288GB of HBM3e memory per GPU, 2.3TB in total, Akamai says. It is preparing its data centres for Nvidia’s Vera Rubin platform in 2027; earlier this month it signed what EDP called its first Asia-Pacific renewable energy agreement, for rooftop solar in Singapore.

“We have been implementing an AI infrastructure strategy based on the notion that AI inference will represent the majority of enterprise focus and investment. Our customers proved that thesis and pulled us further,” said Adam Karon, Akamai’s chief operating officer and general manager of Akamai Cloud.

Akamai is selling the boundary as much as the GPUs

Neither half of this launch is new on its own. Managed model hosting behind an OpenAI-style API is now standard, and the open-weight models in Akamai’s catalogue can be run elsewhere. What Akamai is offering is a package: the model, the place the agent runs, the box its code runs in and the security tools around all three, from one supplier with a very large network.

The sandbox is the more interesting part. Blocking outbound access by default answers the most obvious fear about agents, that code written on the fly will send data somewhere it should not. But WebAssembly and microVMs make different trade-offs: one is light and starts quickly, the other gives code a full operating system to work in. By supporting both, and having helped Perplexity launch SPACE, Akamai is hedging rather than picking a side.

Buyers should read the labels carefully. The release says “beta”, the product page says “tech preview”, the documentation says “closed beta”, and the one-minute step limit rules out long-running agents for now. The $14 billion figure is Akamai’s own and covers its whole cloud business, not these products. The test is whether the promised controls on “what agents can reach” arrive, and whether Akamai publishes prices that let customers compare.